Trust
Trust in Zephr is enforced by three mechanisms working together.
Admission scope
Tool calls that pass a worktreeId outside the admitted scope are rejected with
scope_violation. The trusted launcher (project approve → path bind → launch) establishes
exactly one admitted worktree before any tool call is allowed.
Trust Firewall (zephr_rules)
Per-worktree policy on review state, contradiction handling, content bounds, and recall
eligibility. The Firewall can only ever narrow retrieval, never broaden it — its
recallEligibleStates narrows the fixed core-eligible set (confirmed + valid).
Version-chain integrity
The current-version pointer, contiguous revisions, and non-overlapping transaction intervals
are verified, not assumed, by zephr_verify. Only well-formed, provenanced, confirmed beliefs
are reported verified.
Contradiction and supersession
Beliefs that contradict or supersede one another are tracked and surfaced with directionality, so recall never silently hides a conflict.